A hybrid deep computation model for feature learning on aero-engine data: applications to fault detection

doi:10.1016/j.apm.2020.02.002

Applied Mathematical Modelling

Volume 83, July 2020, Pages 487-496

https://doi.org/10.1016/j.apm.2020.02.002 Get rights and content

Highlights

•
A data-driven hybrid deep computation model is designed to effectively detect gas-path faults in aero-engines.
•
A convolutional and recurrent computing module is designed to model temporal and spatial features of the gas-path data.
•
A convolutional and recurrent back-propagation algorithm is introduced to train the proposed module.
•
Extensive experiments on real aero-engine data are conducted, and illustrate the effectiveness of the hybrid deep computation model.

Abstract

Recently, the safety of aircraft has attracted much attention with some crashes occurring. Gas-path faults, as the most common faults of aircraft, pose a vast challenge for the safety of aircraft because of the complexity of the aero-engine structure. In this article, a hybrid deep computation model is proposed to effectively detect gas-path faults on the basis of the performance data. In detail, to capture the local spatial features of the gas-path performance data, an unfully connected convolutional neural network of one-dimensional kernels is used. Furthermore, to model the temporal patterns hidden in the gas-path faults, a recurrent computation architecture is introduced. Finally, extensive experiments are conducted on real aero-engine data. The results show that the proposed model can outperform the models with which it is compared.

Introduction

With the development of technology, the lives of people are changing greatly. Aircraft are becoming the prevalent vehicle. As a result, the safety of aircraft is attracting much attention. An aero engine is the power source, and its working state has a great effect on the safety of aircraft. Many terrible accidents caused by aero-engine faults have happened. For instance, a Southwest Airlines flight from New York to Dallas crashed because of engine failure [1]. An aero engine can experience many unexpected faults, such as mechanical breakdown and gas-path faults, since it always works in an environment of high temperature, high pressure, and high vibration. Among all aero-engine faults, a gas-path fault is the most common fault. The effective detection of a gas-path fault can prevent a great number of aircraft disasters. Thus, the detection of a gas-path fault is vital research.

In recent years, many detection methods have been proposed to detect gas-path faults. For example, Xue and Wang [2] proposed a fusion method based on Kalman filters to detect gas-path faults. Sun et al. [3] designed a set of rules that can combine conflicting evidence based on rough set and Dempster–Shafer theories. These mathematical-model-based methods cannot model the intrinsic fault patterns well because of the complexity of the aero-engine structure. To solve this problem, some researchers devised data-driven methods to detect gas-path faults. For example, Yu and Huang [4] used a least-squares support vector machine (SVM) to classify aero-engine faults. Zhao and Huang [5] designed a neural network to learn features of the diagnosis data. Those data-driven methods can increase the accuracy of detection of gas-path faults. However, they are shallow models that cannot capture intrinsic fault patterns of more advanced modern aero engines. Thus, the accurate detection of gas-path faults requires novel methods.

Deep learning, as a novel method, can well perceive high-order state features of big data and produce effective nonlinear representations [6], [7]. It has been widely used in image classification [8] and speech recognition [9]. Recently, inspired by outstanding results of deep learning in other domains, the application of deep learning to gas-path fault detection in aero engines has attracted much attention.

In this article, a hybrid deep computation (HDC) model is proposed for the detection of aero-engine gas-path faults. In detail, an unfully connected convolutional neural network (CNN) of one-dimensional kernels is used to capture the local features of the gas-path performance data, which can take advantage of the space topologies hidden in the gas-path data. Furthermore, a recurrent computation architecture is used to mine the temporal dependencies of fault patterns in the gas-path data. Then, on the basis of the two aforementioned computing modules, a convolutional and recurrent architecture is proposed to capture the instinct patterns of the gas-path faults. Finally, to evaluate the HDC model, extensive experiments are conducted on real aero-engine data. The results show that the proposed model can outperform the models with which it is compared. Thus, the contributions of this work can be summarized as follows:

•
To improve the safety of aircraft, an HDC model is designed that can effectively detect gas-path faults of aero engines.
•
To capture instinct patterns of gas-path faults, convolutional and recurrent computing modules are designed that can learn the spatial and temporal features of gas-path data. Furthermore, the corresponding back-propagation rules are introduced to train the proposed modules.
•
To evaluate the proposed model, extensive experiments are conducted on real aero-engine data, and illustrate the effectiveness of the HDC model.

The rest of this article is organized as follows. In Section 2, some related studies are briefly reviewed. The details of the proposed model are given in Section 3, and the hybrid error back-propagation is introduced in Section 4. In Section 5, the detailed experimental setting and results are provided. Finally, Section 6 concludes the article.

Section snippets

Related work

In this article, a convolutional and recurrent architecture is proposed to detect the patterns of faults in gas-path data sequences. Thus, related studies on fault detection, CNNs, and recurrent neural networks (RNNs) are introduced in this section.

The HDC model

In this section, an HDC model is proposed to learn intrinsic features of aero-engine data for gas-path fault detection. In particular, each gas-path data collected continuously from performance parameters of aero engines is first split into k time steps as the input. Then, the spatial and temporal features are modeled in a layerwise manner. The HDC network is composed of a one-dimensional convolutional network, a multilayer perceptron (MLP), and a recurrent network, as shown in Fig. 1. In the

HDC learning for fault detection

As shown in Fig. 1, each gas-path parameter sequential data modeled as a data block is fed into HDC. Then, the output feature serves as input into a softmax layer to recognize the gas-path parameter sequential data. The HDC learning is based on the back-propagation algorithm optimized by minimizing the cross-entropy loss. It is composed of four stages: output layer learning, LSTM network learning, MLP network learning, and CNN learning.

Experiments

In this section, we report extensive experiments on real aero-engine data to evaluate the HDC model. All the experiments were performed with a server with 64-GB memory and a 10-core, 20-thread Intel Xeon E7-4800 CPU.

Conclusion

In this article, an HDC model was proposed for aero-engine gas-path fault detection. The HDC model is composed of a convolution network, an MLP network, and a recurrent network. It learns the spatial and temporal dependencies hidden in the gas-path data of aero engines. Extensive experimental results show that HDC outperforms other common methods in the fault detection task for aero engines and is more robust to noise.

Acknowledgments

This work was supported in part by the Key Laboratory of Vibration and Control of Aero-Propulsion Systems of the Ministry of Education, Northeastern University, under grant no. VCAME201705, the National Natural Science Foundation of China under grants no. 61602083 and no. 61672123, the Fundamental Research Funds for the Central Universities under grant no. DUT2017TB0, the Dalian University of Technology Fundamental Research Fund under grant no. DUT15RC(3)100, and the Doctoral Scientific

References (28)

Q. Zhang et al.
A survey on deep learning for big data
Inf. Fusion
(2018)
J. Gao et al.
A canonical polyadic deep convolutional computation model for big data feature learning in internet of things
Future Gener. Comput. Syst.
(2019)
2018. https://en.wikipedia.org/wiki/Southwest_Airlines_Flight_1380 (accessed 25 August...
W. Xue et al.
Application of a bank of Kalman filters for aircraft engine sensor actuator fault diagnosis
Comput. Measur. Control
(2015)
W. Sun et al.
Avionic devices fault diagnosis based on fusion method of rough set and d-s theory
J. Beijing Univ. Aeronaut. Astronaut.
(2015)
G. Yu et al.
Aero-engine fault diagnosis based on support vector machine and Kalman filter
Aeroengine
(2012)
S. Zhao et al.
Fault diagnosis for Aeroengine gas path components based on neural network multisensory data fusion
J. Aerosp. Power
(2008)
P. Li et al.
An improved stacked auto-encoder for network traffic flow classification
IEEE Netw.
(2018)
P. Li et al.
An incremental deep convolutional computation model for feature learning on industrial big data
IEEE Trans. Ind. Informat.
(2018)
F. Lu et al.
Application of multi-outputs LSSVR by PSO to the aero-engine model
J. Syst. Eng. Electr.
(2009)

Q. Xu et al.

Aero-engine fault diagnosis by a new fast multi-class support vector algorithm

J. Propul. Technol.

(2012)

Q. Huang et al.

A kind of approach for aero engine gas path fault diagnosis

Proceedings of the IEEE International Conference on Prognostics and Health Management, Dallas, TX, USA

(2017)

Y. Pan et al.

Aero-engine gas path fault diagnosis based on broyden algorithm

Journal of Propulsion Technology

(2017)

A. Krizhevsky et al.

Imagenet classification with deep convolutional neural networks

Proceedings of the Advances in Neural Information Processing Systems 25, Lake Tahoe, Nevada, United States

(2012)

Cited by (12)

A variational level set model based on local-global function approximation for image segmentation
2024, Digital Signal Processing: A Review Journal
The variational level set model is widely used in image segmentation tasks due to its sound theoretical basis and reliable performance. However, it often faces limitations when dealing with images that contain intensity inhomogeneity and noise. To address this challenge, a novel variational level set model based on local-global function approximation (LSM-LGA) is proposed in this paper. In this approach, an approximation function space is first introduced for the input image. This space defines the function as a linear combination of a local approximation image and a global one, weighted by two coefficients and two complementary characteristic functions defined by a level set function. To achieve the optimal representation of the input image within this approximation function space, a hybrid-metric strategy that combines Euclidean distance and Jeffreys divergence is introduced. An alternating direction method of multiplier (ADMM) based on Euclidean distance is developed to solve for the optimal weight coefficients of the local and global approximation images. Additionally, gradient descent based on Jeffreys divergence is utilized to solve for the optimal level set function. Furthermore, the existence and uniqueness of the optimal approximation function are proven using the theory of projection onto convex sets. Extensive numerical experiments on natural and synthetic images with intensity inhomogeneity, textures, and noise validate that the LSM-LGA exhibits superior performance compared with several state-of-the-art models in terms of segmentation quality and computational efficiency.
Multi-attribute Bayesian fault prediction for hidden-state systems under condition monitoring
2022, Applied Mathematical Modelling
Citation Excerpt :
For hidden-state systems, the partial observations of health status may be multivariate which contains dependence information of deterioration [19]. Current research works analyzed the multivariate observations by modeling each signal separately or by using dimension-reduction techniques, which may ignore some valuable information for degradation modeling [20]. In this paper, our work is to present a multivariate stochastic model in which the multidimensional data can be fully exploited.
Although Bayesian approaches have been utilized in engineering systems for health prognostics, very little work has been done using Bayesian methods for fault prediction of systems under multiple attributes. To address this issue, in this paper a novel multi-attribute Bayesian control chart is presented for predicting failures of hidden-state systems by jointly considering two performance measures of system operation. The system actual status is represented by a three-state multivariate hidden stochastic process with a normal state, an abnormal state, and a failure state. The working states are unobservable and failure state is observable. Based on the built hidden-state model, a fault prediction scheme integrating both system availability and cost objectives is constructed via a multi-attribute Bayesian control chart to monitor and predict impending risks of the operational systems. The Bayesian control chart alarms when the probability of impending risks reaches a certain control limit, which is optimized and determined by a computational algorithm developed in a semi-Markov decision process framework. The proposed fault prediction scheme provides an appearing feature to jointly consider multiple attributes for hidden-state systems. A real case study of mechanical generators is presented and a comparison with other Bayesian and non-Bayesian methods is also given, which demonstrates the effectiveness and superiority of the proposed approach.
Unilateral alignment transfer neural network for fault diagnosis of aircraft engine
2021, Aerospace Science and Technology
Citation Excerpt :
By this means, a great deal of manual labor and unplanned downtime can be avoided, and the maintenance cycle is shortened. Hence, intelligent fault diagnosis has become one of the hot topic in various engineering applications, such as aerospace facilities [1–6], aircraft engines [7–13], rotating machinery [14–18], and electrical vehicles [19,20]. In recent years, condition-based maintenance [21–23], which is developed to accomplish reliable inference of the actual state of in service aircraft engines and design corresponding maintenance strategy, is crucial to gas turbine industry, because most aircraft engines always operate under extreme working conditions, such as intolerable high temperature, pressure and loads.
Data-driven fault diagnosis for aircraft engine has recently been widely studied, most of which utilized traditional machine learning tools. However, these approaches often assumed that the data used for training and testing come from the identical distribution, which is unpractical in practice, because even for the same type of engine, there may be some discrepancies in the data collected between different engine individuals. This idealized assumption may confine these promising data-driven technologies to well-design experimental environments instead of applying them to practical applications. Meanwhile, the research on data-driven engine fault diagnosis methods based on transfer learning is relatively limited. To overcome this issue, this paper proposes a methodology using unilateral alignment strategy to address the issue of cross domain engine gas path fault diagnosis, where the training data extracted from source domain and the testing data extracted from target domain are assumed to come from different distributions. More importantly, the proposed methodology is capable of maintaining the inter-class relationship of source domain, instead of aligning the discrepancies for source and target domain by mapping both of their feature spaces into an unknown intermediate space, because the label space between two domains may be different, and the label set of target domain might be only a subset of that of source domain. Meanwhile, both the marginal and conditional distribution discrepancies are considered. Finally, the proposed approaches are evaluated by extensive experiments on engine cross domain fault diagnosis, including hybrid transfer experiments, complete transfer experiments and missing class experiments, and the overall results demonstrate the feasibility of the proposed methodology.
Gas turbine aero engine fault detection using Geo-TLSVM and digital twin with multimodal data analysis
2024, Engineering Research Express
A Variational Level Set Model Based on Local-Global Function Approximation for Image Segmentation
2023, SSRN
A deep learning-based prognostic framework for aeroengine exhaust gas temperature margin
2023, Revista Internacional de Metodos Numericos para Calculo y Diseno en Ingenieria

View all citing articles on Scopus

View full text

A hybrid deep computation model for feature learning on aero-engine data: applications to fault detection

Highlights

Abstract

Introduction

Section snippets

Related work

The HDC model

HDC learning for fault detection

Experiments

Conclusion

Acknowledgments

Inf. Fusion

Future Gener. Comput. Syst.

Application of a bank of Kalman filters for aircraft engine sensor actuator fault diagnosis

Comput. Measur. Control

Avionic devices fault diagnosis based on fusion method of rough set and d-s theory

J. Beijing Univ. Aeronaut. Astronaut.

Aero-engine fault diagnosis based on support vector machine and Kalman filter

Aeroengine

Fault diagnosis for Aeroengine gas path components based on neural network multisensory data fusion

J. Aerosp. Power

An improved stacked auto-encoder for network traffic flow classification

IEEE Netw.

An incremental deep convolutional computation model for feature learning on industrial big data

IEEE Trans. Ind. Informat.

Application of multi-outputs LSSVR by PSO to the aero-engine model

J. Syst. Eng. Electr.